libs/wine/mmap.c

   1 /*
   2  * Wine memory mappings support
   3  *
   4  * Copyright 2000, 2004 Alexandre Julliard
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  19  */
  20
  21 #include "config.h"
  22 #include "wine/port.h"
  23
  24 #include <assert.h>
  25 #include <ctype.h>
  26 #include <fcntl.h>
  27 #include <stdlib.h>
  28 #include <stdio.h>
  29 #include <string.h>
  30 #include <sys/types.h>
  31 #ifdef HAVE_SYS_MMAN_H
  32 #include <sys/mman.h>
  33 #endif
  34 #ifdef HAVE_UNISTD_H
  35 # include <unistd.h>
  36 #endif
  37 #ifdef HAVE_STDINT_H
  38 # include <stdint.h>
  39 #endif
  40
  41 #include "wine/library.h"
  42 #include "wine/list.h"
  43
  44 #ifdef HAVE_MMAP
  45
  46 struct reserved_area
  47 {
  48     struct list entry;
  49     void       *base;
  50     size_t      size;
  51 };
  52
  53 static struct list reserved_areas = LIST_INIT(reserved_areas);
  54 static const unsigned int granularity_mask = 0xffff;  /* reserved areas have 64k granularity */
  55
  56 #ifndef MAP_NORESERVE
  57 #define MAP_NORESERVE 0
  58 #endif
  59 #ifndef MAP_PRIVATE
  60 #define MAP_PRIVATE 0
  61 #endif
  62 #ifndef MAP_ANON
  63 #define MAP_ANON 0
  64 #endif
  65
  66 static inline int get_fdzero(void)
  67 {
  68     static int fd = -1;
  69
  70     if (MAP_ANON == 0 && fd == -1)
  71     {
  72         if ((fd = open( "/dev/zero", O_RDONLY )) == -1)
  73         {
  74             perror( "/dev/zero: open" );
  75             exit(1);
  76         }
  77     }
  78     return fd;
  79 }
  80
  81 #if (defined(__svr4__) || defined(__NetBSD__)) && !defined(MAP_TRYFIXED)
  82 /***********************************************************************
  83  *             try_mmap_fixed
  84  *
  85  * The purpose of this routine is to emulate the behaviour of
  86  * the Linux mmap() routine if a non-NULL address is passed,
  87  * but the MAP_FIXED flag is not set.  Linux in this case tries
  88  * to place the mapping at the specified address, *unless* the
  89  * range is already in use.  Solaris, however, completely ignores
  90  * the address argument in this case.
  91  *
  92  * As Wine code occasionally relies on the Linux behaviour, e.g. to
  93  * be able to map non-relocatable PE executables to their proper
  94  * start addresses, or to map the DOS memory to 0, this routine
  95  * emulates the Linux behaviour by checking whether the desired
  96  * address range is still available, and placing the mapping there
  97  * using MAP_FIXED if so.
  98  */
  99 static int try_mmap_fixed (void *addr, size_t len, int prot, int flags,
 100                            int fildes, off_t off)
 101 {
 102     char * volatile result = NULL;
 103     int pagesize = getpagesize();
 104     pid_t pid;
 105
 106     /* We only try to map to a fixed address if
 107        addr is non-NULL and properly aligned,
 108        and MAP_FIXED isn't already specified. */
 109
 110     if ( !addr )
 111         return 0;
 112     if ( (uintptr_t)addr & (pagesize-1) )
 113         return 0;
 114     if ( flags & MAP_FIXED )
 115         return 0;
 116
 117     /* We use vfork() to freeze all threads of the
 118        current process.  This allows us to check without
 119        race condition whether the desired memory range is
 120        already in use.  Note that because vfork() shares
 121        the address spaces between parent and child, we
 122        can actually perform the mapping in the child. */
 123
 124     if ( (pid = vfork()) == -1 )
 125     {
 126         perror("try_mmap_fixed: vfork");
 127         exit(1);
 128     }
 129     if ( pid == 0 )
 130     {
 131         int i;
 132         char vec;
 133
 134         /* We call mincore() for every page in the desired range.
 135            If any of these calls succeeds, the page is already
 136            mapped and we must fail. */
 137         for ( i = 0; i < len; i += pagesize )
 138             if ( mincore( (caddr_t)addr + i, pagesize, &vec ) != -1 )
 139                _exit(1);
 140
 141         /* Perform the mapping with MAP_FIXED set.  This is safe
 142            now, as none of the pages is currently in use. */
 143         result = mmap( addr, len, prot, flags | MAP_FIXED, fildes, off );
 144         if ( result == addr )
 145             _exit(0);
 146
 147         if ( result != (void *) -1 ) /* This should never happen ... */
 148             munmap( result, len );
 149
 150        _exit(1);
 151     }
 152
 153     /* vfork() lets the parent continue only after the child
 154        has exited.  Furthermore, Wine sets SIGCHLD to SIG_IGN,
 155        so we don't need to wait for the child. */
 156
 157     return result == addr;
 158 }
 159
 160 #elif defined(__APPLE__)
 161
 162 #include <mach/mach_init.h>
 163 #include <mach/vm_map.h>
 164
 165 /*
 166  * On Darwin, we can use the Mach call vm_allocate to allocate
 167  * anonymous memory at the specified address, and then use mmap with
 168  * MAP_FIXED to replace the mapping.
 169  */
 170 static int try_mmap_fixed (void *addr, size_t len, int prot, int flags,
 171                            int fildes, off_t off)
 172 {
 173     vm_address_t result = (vm_address_t)addr;
 174
 175     if (!vm_allocate(mach_task_self(),&result,len,0))
 176     {
 177         if (mmap( (void *)result, len, prot, flags | MAP_FIXED, fildes, off ) != MAP_FAILED)
 178             return 1;
 179         vm_deallocate(mach_task_self(),result,len);
 180     }
 181     return 0;
 182 }
 183
 184 #endif  /* (__svr4__ || __NetBSD__) && !MAP_TRYFIXED */
 185
 186
 187 /***********************************************************************
 188  *              wine_anon_mmap
 189  *
 190  * Portable wrapper for anonymous mmaps
 191  */
 192 void *wine_anon_mmap( void *start, size_t size, int prot, int flags )
 193 {
 194 #ifdef MAP_SHARED
 195     flags &= ~MAP_SHARED;
 196 #endif
 197
 198     /* Linux EINVAL's on us if we don't pass MAP_PRIVATE to an anon mmap */
 199     flags |= MAP_PRIVATE | MAP_ANON;
 200
 201     if (!(flags & MAP_FIXED))
 202     {
 203 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 204         /* Even FreeBSD 5.3 does not properly support NULL here. */
 205         if( start == NULL ) start = (void *)0x110000;
 206 #endif
 207
 208 #ifdef MAP_TRYFIXED
 209         /* If available, this will attempt a fixed mapping in-kernel */
 210         flags |= MAP_TRYFIXED;
 211 #elif defined(__svr4__) || defined(__NetBSD__) || defined(__APPLE__)
 212         if ( try_mmap_fixed( start, size, prot, flags, get_fdzero(), 0 ) )
 213             return start;
 214 #endif
 215     }
 216     return mmap( start, size, prot, flags, get_fdzero(), 0 );
 217 }
 218
 219
 220 /***********************************************************************
 221  *              mmap_reserve
 222  *
 223  * mmap wrapper used for reservations, only maps the specified address
 224  */
 225 static inline int mmap_reserve( void *addr, size_t size )
 226 {
 227     void *ptr;
 228     int flags = MAP_PRIVATE | MAP_ANON | MAP_NORESERVE;
 229
 230 #ifdef MAP_TRYFIXED
 231     flags |= MAP_TRYFIXED;
 232 #elif defined(__APPLE__)
 233     return try_mmap_fixed( addr, size, PROT_NONE, flags, get_fdzero(), 0 );
 234 #endif
 235     ptr = mmap( addr, size, PROT_NONE, flags, get_fdzero(), 0 );
 236     if (ptr != addr && ptr != (void *)-1)  munmap( ptr, size );
 237     return (ptr == addr);
 238 }
 239
 240
 241 /***********************************************************************
 242  *           reserve_area
 243  *
 244  * Reserve as much memory as possible in the given area.
 245  */
 246 #ifdef __i386__
 247 static void reserve_area( void *addr, void *end )
 248 {
 249     size_t size = (char *)end - (char *)addr;
 250
 251 #if (defined(__svr4__) || defined(__NetBSD__)) && !defined(MAP_TRYFIXED)
 252     /* try_mmap_fixed is inefficient when using vfork, so we need a different algorithm here */
 253     /* we assume no other thread is running at this point */
 254     size_t i, pagesize = getpagesize();
 255     char vec;
 256
 257     while (size)
 258     {
 259         for (i = 0; i < size; i += pagesize)
 260             if (mincore( (caddr_t)addr + i, pagesize, &vec ) != -1) break;
 261
 262         i &= ~granularity_mask;
 263         if (i && mmap( addr, i, PROT_NONE, MAP_FIXED | MAP_PRIVATE | MAP_ANON | MAP_NORESERVE,
 264                        get_fdzero(), 0 ) != (void *)-1)
 265             wine_mmap_add_reserved_area( addr, i );
 266
 267         i += granularity_mask + 1;
 268         if ((char *)addr + i < (char *)addr) break;  /* overflow */
 269         addr = (char *)addr + i;
 270         if (addr >= end) break;
 271         size = (char *)end - (char *)addr;
 272     }
 273 #else
 274     if (!size) return;
 275
 276     if (mmap_reserve( addr, size ))
 277     {
 278         wine_mmap_add_reserved_area( addr, size );
 279         return;
 280     }
 281     if (size > granularity_mask + 1)
 282     {
 283         size_t new_size = (size / 2) & ~granularity_mask;
 284         reserve_area( addr, (char *)addr + new_size );
 285         reserve_area( (char *)addr + new_size, end );
 286     }
 287 #endif
 288 }
 289
 290
 291 /***********************************************************************
 292  *           reserve_malloc_space
 293  *
 294  * Solaris malloc is not smart enough to obtain space through mmap(), so try to make
 295  * sure that there is some available sbrk() space before we reserve other things.
 296  */
 297 static void reserve_malloc_space( size_t size )
 298 {
 299 #ifdef __sun
 300     size_t i, count = size / 1024;
 301     void **ptrs = malloc( count * sizeof(ptrs[0]) );
 302
 303     if (!ptrs) return;
 304
 305     for (i = 0; i < count; i++) if (!(ptrs[i] = malloc( 1024 ))) break;
 306     if (i--)  /* free everything except the last one */
 307         while (i) free( ptrs[--i] );
 308     free( ptrs );
 309 #endif
 310 }
 311
 312 #endif  /* __i386__ */
 313
 314
 315 /***********************************************************************
 316  *           reserve_dos_area
 317  *
 318  * Reserve the DOS area (0x00000000-0x00110000).
 319  */
 320 static void reserve_dos_area(void)
 321 {
 322     const size_t page_size = getpagesize();
 323     const size_t dos_area_size = 0x110000;
 324     void *ptr;
 325
 326     /* first page has to be handled specially */
 327     ptr = wine_anon_mmap( (void *)page_size, dos_area_size - page_size, PROT_NONE, MAP_NORESERVE );
 328     if (ptr != (void *)page_size)
 329     {
 330         if (ptr != (void *)-1) munmap( ptr, dos_area_size - page_size );
 331         return;
 332     }
 333     /* now add first page with MAP_FIXED */
 334     wine_anon_mmap( NULL, page_size, PROT_NONE, MAP_NORESERVE|MAP_FIXED );
 335     wine_mmap_add_reserved_area( NULL, dos_area_size );
 336 }
 337
 338
 339 /***********************************************************************
 340  *           mmap_init
 341  */
 342 void mmap_init(void)
 343 {
 344     struct reserved_area *area;
 345     struct list *ptr;
 346 #ifdef __i386__
 347     char stack;
 348     char * const stack_ptr = &stack;
 349     char *user_space_limit = (char *)0x7ffe0000;
 350
 351     reserve_malloc_space( 8 * 1024 * 1024 );
 352
 353     /* check for a reserved area starting at the user space limit */
 354     /* to avoid wasting time trying to allocate it again */
 355     LIST_FOR_EACH( ptr, &reserved_areas )
 356     {
 357         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 358         if ((char *)area->base > user_space_limit) break;
 359         if ((char *)area->base + area->size > user_space_limit)
 360         {
 361             user_space_limit = (char *)area->base + area->size;
 362             break;
 363         }
 364     }
 365
 366     if (stack_ptr >= user_space_limit)
 367     {
 368         char *end = 0;
 369         char *base = stack_ptr - ((unsigned int)stack_ptr & granularity_mask) - (granularity_mask + 1);
 370         if (base > user_space_limit) reserve_area( user_space_limit, base );
 371         base = stack_ptr - ((unsigned int)stack_ptr & granularity_mask) + (granularity_mask + 1);
 372 #if defined(linux) || defined(__FreeBSD__)
 373         /* Heuristic: assume the stack is near the end of the address */
 374         /* space, this avoids a lot of futile allocation attempts */
 375         end = (char *)(((unsigned long)base + 0x0fffffff) & 0xf0000000);
 376 #endif
 377         reserve_area( base, end );
 378     }
 379     else reserve_area( user_space_limit, 0 );
 380 #endif /* __i386__ */
 381
 382     /* reserve the DOS area if not already done */
 383
 384     ptr = list_head( &reserved_areas );
 385     if (ptr)
 386     {
 387         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 388         if (!area->base) return;  /* already reserved */
 389     }
 390     reserve_dos_area();
 391 }
 392
 393
 394 /***********************************************************************
 395  *           wine_mmap_add_reserved_area
 396  *
 397  * Add an address range to the list of reserved areas.
 398  * Caller must have made sure the range is not used by anything else.
 399  *
 400  * Note: the reserved areas functions are not reentrant, caller is
 401  * responsible for proper locking.
 402  */
 403 void wine_mmap_add_reserved_area( void *addr, size_t size )
 404 {
 405     struct reserved_area *area;
 406     struct list *ptr;
 407
 408     if (!((char *)addr + size)) size--;  /* avoid wrap-around */
 409
 410     LIST_FOR_EACH( ptr, &reserved_areas )
 411     {
 412         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 413         if (area->base > addr)
 414         {
 415             /* try to merge with the next one */
 416             if ((char *)addr + size == (char *)area->base)
 417             {
 418                 area->base = addr;
 419                 area->size += size;
 420                 return;
 421             }
 422             break;
 423         }
 424         else if ((char *)area->base + area->size == (char *)addr)
 425         {
 426             /* merge with the previous one */
 427             area->size += size;
 428
 429             /* try to merge with the next one too */
 430             if ((ptr = list_next( &reserved_areas, ptr )))
 431             {
 432                 struct reserved_area *next = LIST_ENTRY( ptr, struct reserved_area, entry );
 433                 if ((char *)addr + size == (char *)next->base)
 434                 {
 435                     area->size += next->size;
 436                     list_remove( &next->entry );
 437                     free( next );
 438                 }
 439             }
 440             return;
 441         }
 442     }
 443
 444     if ((area = malloc( sizeof(*area) )))
 445     {
 446         area->base = addr;
 447         area->size = size;
 448         list_add_before( ptr, &area->entry );
 449     }
 450 }
 451
 452
 453 /***********************************************************************
 454  *           wine_mmap_remove_reserved_area
 455  *
 456  * Remove an address range from the list of reserved areas.
 457  * If 'unmap' is non-zero the range is unmapped too.
 458  *
 459  * Note: the reserved areas functions are not reentrant, caller is
 460  * responsible for proper locking.
 461  */
 462 void wine_mmap_remove_reserved_area( void *addr, size_t size, int unmap )
 463 {
 464     struct reserved_area *area;
 465     struct list *ptr;
 466
 467     if (!((char *)addr + size)) size--;  /* avoid wrap-around */
 468
 469     ptr = list_head( &reserved_areas );
 470     /* find the first area covering address */
 471     while (ptr)
 472     {
 473         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 474         if ((char *)area->base >= (char *)addr + size) break;  /* outside the range */
 475         if ((char *)area->base + area->size > (char *)addr)  /* overlaps range */
 476         {
 477             if (area->base >= addr)
 478             {
 479                 if ((char *)area->base + area->size > (char *)addr + size)
 480                 {
 481                     /* range overlaps beginning of area only -> shrink area */
 482                     if (unmap) munmap( area->base, (char *)addr + size - (char *)area->base );
 483                     area->size -= (char *)addr + size - (char *)area->base;
 484                     area->base = (char *)addr + size;
 485                     break;
 486                 }
 487                 else
 488                 {
 489                     /* range contains the whole area -> remove area completely */
 490                     ptr = list_next( &reserved_areas, ptr );
 491                     if (unmap) munmap( area->base, area->size );
 492                     list_remove( &area->entry );
 493                     free( area );
 494                     continue;
 495                 }
 496             }
 497             else
 498             {
 499                 if ((char *)area->base + area->size > (char *)addr + size)
 500                 {
 501                     /* range is in the middle of area -> split area in two */
 502                     struct reserved_area *new_area = malloc( sizeof(*new_area) );
 503                     if (new_area)
 504                     {
 505                         new_area->base = (char *)addr + size;
 506                         new_area->size = (char *)area->base + area->size - (char *)new_area->base;
 507                         list_add_after( ptr, &new_area->entry );
 508                     }
 509                     else size = (char *)area->base + area->size - (char *)addr;
 510                     area->size = (char *)addr - (char *)area->base;
 511                     if (unmap) munmap( addr, size );
 512                     break;
 513                 }
 514                 else
 515                 {
 516                     /* range overlaps end of area only -> shrink area */
 517                     if (unmap) munmap( addr, (char *)area->base + area->size - (char *)addr );
 518                     area->size = (char *)addr - (char *)area->base;
 519                 }
 520             }
 521         }
 522         ptr = list_next( &reserved_areas, ptr );
 523     }
 524 }
 525
 526
 527 /***********************************************************************
 528  *           wine_mmap_is_in_reserved_area
 529  *
 530  * Check if the specified range is included in a reserved area.
 531  * Returns 1 if range is fully included, 0 if range is not included
 532  * at all, and -1 if it is only partially included.
 533  *
 534  * Note: the reserved areas functions are not reentrant, caller is
 535  * responsible for proper locking.
 536  */
 537 int wine_mmap_is_in_reserved_area( void *addr, size_t size )
 538 {
 539     struct reserved_area *area;
 540     struct list *ptr;
 541
 542     LIST_FOR_EACH( ptr, &reserved_areas )
 543     {
 544         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 545         if (area->base > addr) break;
 546         if ((char *)area->base + area->size <= (char *)addr) continue;
 547         /* area must contain block completely */
 548         if ((char *)area->base + area->size < (char *)addr + size) return -1;
 549         return 1;
 550     }
 551     return 0;
 552 }
 553
 554
 555 /***********************************************************************
 556  *           wine_mmap_enum_reserved_areas
 557  *
 558  * Enumerate the list of reserved areas, sorted by addresses.
 559  * If enum_func returns a non-zero value, enumeration is stopped and the value is returned.
 560  *
 561  * Note: the reserved areas functions are not reentrant, caller is
 562  * responsible for proper locking.
 563  */
 564 int wine_mmap_enum_reserved_areas( int (*enum_func)(void *base, size_t size, void *arg), void *arg,
 565                                    int top_down )
 566 {
 567     int ret = 0;
 568     struct list *ptr;
 569
 570     if (top_down)
 571     {
 572         for (ptr = reserved_areas.prev; ptr != &reserved_areas; ptr = ptr->prev)
 573         {
 574             struct reserved_area *area = LIST_ENTRY( ptr, struct reserved_area, entry );
 575             if ((ret = enum_func( area->base, area->size, arg ))) break;
 576         }
 577     }
 578     else
 579     {
 580         for (ptr = reserved_areas.next; ptr != &reserved_areas; ptr = ptr->next)
 581         {
 582             struct reserved_area *area = LIST_ENTRY( ptr, struct reserved_area, entry );
 583             if ((ret = enum_func( area->base, area->size, arg ))) break;
 584         }
 585     }
 586     return ret;
 587 }
 588
 589 #else /* HAVE_MMAP */
 590
 591 void mmap_init(void)
 592 {
 593 }
 594
 595 #endif